1st Supervisor: Alistair Young, King’s College London
2nd Supervisor: Claudia Prieto, King’s College London
Additional Supervisor: Rene Botnar, King’s College London
Clinical Champion: Amedeo Chiribiri, King’s College London
Aim of the PhD Project:
- Train a machine learning algorithm to characterise heart shape and scar burden in 3D multi-contrast magnetic resonance images of the heart
- Develop methods to adapt machine learning applications to new imaging methods.
- Determine the clinical efficacy of fully automated analysis compared to current methods.
Project Description / Background:
Cardiovascular disease is the primary cause of death world-wide. Magnetic resonance imaging can identify injury within the heart muscle in many types of disease, including coronary artery disease, cardiomyopathies, and atrial fibrillation. Conventional 2D inversion recovery (IR) and phase sensitive inversion recovery (PSIR) late gadolinium enhancement (LGE) cardiovascular magnetic resonance (CMR) imaging is widely used clinically for the assessment of myocardial viability and disease burden. However, artefacts due to fat, poor spatial resolution, lack of contrast between scar and blood, and volumetric coverage within a clinically feasible scan time, are currently compromising application in the clinic. In order to address these challenges, at KCL we have recently developed a highly efficient respiratory-motion-corrected 3D whole-heart water/fat LGE imaging framework for both bright and grey blood PSIR imaging [1-2]. In particular, the contrast between scar tissue and blood, resolution, fat separation, and volumetric coverage are all improved with this approach. This enables analysis of scar in the atria, which is very important for patients with atrial fibrillation. However, such improvements to the imaging lead to differences in appearance and dimension (3D vs. 2D, multi-contrast vs. single contrast), so machine learning methods trained on older images often fail when presented with the new images. This is a common problem in machine learning, in which algorithms need to be continuously retrained to adjust to new imaging methods. This is expensive since the training typically requires manual annotation of many thousands of images. This project will develop automatic tools for 3D whole-heart high-resolution LGE image analysis using machine-learning domain adaptation methods [3-4]. Analysis will include automatic myocardial scar segmentation as well as automatic myocardial infarct heterogeneity quantification (infarct core, border zone and healthy myocardium). In particular, methods will focus on the automatic analysis of atrial scar burden. These methods will exploit historical acquisitions of standard 2D bright- blood LGE images (N=100), together with previously performed manual annotations, as well as numerical simulations of the image formation process from in-silico heart models (digital phantoms), to learn features useful for analysis of CMR LGE images. These source images and features will be adapted to the target domain (new imaging methods), using a combination of unsupervised domain adaptation , self-supervised learning  and digital phantom simulations  to discover which can best be utilized in the new, improved 3D free-breathing LGE acquisitions. This will be used to gain an understanding of how to best utilize existing data and simulations to improve robustness and quality in new imaging methods without needing to manually annotate a large number of cases. The clinical efficacy of these methods will be tested in N = 30 patients with suspected ischemic cardiovascular disease prospectively recruited for this study. Also, the proposed automatic LGE image analysis methods will be applied to patients with cardiovascular disease in the clinical workflow at St Thomas’ Hospital. This project would suit an engineer interested in machine learning and image analysis for application in benefiting the treatment of patients with cardiovascular disease.
Figure 1: Example of a patient with scar (arrows) shown as bright signal. The 3D image was acquired in 6 min 28 sec, with resolution of 2mm. The image is resliced into standard views (coronal, short and 2 chamber and 4 chamber). PSIR: phase sensitive inversion recovery.
- Munoz, C., Bustin, A., Neji, R. et al. Motion-corrected 3D whole-heart water-fat high-resolution late gadolinium enhancement cardiovascular magnetic resonance imaging. J Cardiovasc Magn Reson 2010;22:53. doi:10.1186/s12968-020-00649-5
- Milotta G et al ISMRM 2020. 3D Whole-heart Motion Compensated Grey-blood Late Gadolinium Enhancement Imaging.
- Ben-David S et al. A theory of learning from different domains. Machine Learning 2010;79: 151–175. doi: 10.1007/s10994-009-5152-4.
- Yan W et al. The Domain Shift Problem of Medical Image Segmentation and Vendor- Adaptation by Unet-GAN. MICCAI 2019 LNCS 11765 pp. 623–631. doi: 10.1007/978-3-030-32245-8_69
- Chen C et al. Synergistic Image and Feature Adaptation: Towards Cross-Modality Domain Adaptation for Medical Image Segmentation (2018). arXiv:1901.08211v4
- Bai W et al. Self-Supervised Learning for Cardiac MR Image Segmentation by Anatomical Position Prediction. (2019) arXiv:1907.02757v1
- Segars et al. Application of the 4D XCAT Phantoms in Biomedical Imaging and Beyond. IEEE Trans Med Imaging. 2018 37(3): 680–692.